Game Development Reference
In-Depth Information
Regardless of their motives, people playing the game pick numbers. We need to
find a way of modeling those picks. We do not start from the lowest level of
why
they pick the way they do, however. We need to generate the equivalent of what the
descriptive decision theory would show us—that four out of five dentists recom-
mend sugarless gum, for example. If we were to look at the histogram of the guesses
that the players offered, we could identify the populations we mentioned above.
Looking at Figure 11.2 again (which has been grossly simplified from the orig-
inal histogram from Chapter 6), we have identified our three groups. Spread across
the entire spectrum from 0 to 100 are the random guessers. Those are the small
number of people who will guess anything without rhyme or reason. Additionally,
we saw that there were two spikes of people who guessed 33 and 22, respectively. In
Chapter 6 we showed why people would have guessed those particular logical an-
swers. The darkened section in the figure represents our mystery population. There
was an obvious bulge in the lower half of the histogram. Many people guessed those
semi-logical answers—far more so than guessed randomly. Unlike the random
guessers and the targeted 33/22 folks, we have no idea what these people were
thinking. We only know that they tended to be between approximately 10 and 50,
with a distinct bulge between 20 and 30.
Re-Creating the Bulge
Turning to artificial intelligence (AI), if we wanted to model an agent that would
play the game
the way humans do
, we would have to take
all
of these populations
into account. Creating an agent that generated purely random guesses wouldn't
create the histogram that the Denmark study yielded. Creating an agent that logi-
cally solved the answer would generate the same answer every time—entirely de-
pendant on the single logical formula we devised. Neither of these solutions, when
repeated enough times to resemble a large population of people would mimic the
results that 19,000
very real
people gave us in the Denmark study.
The solution is to incorporate three approaches. If we were to assume that a
small portion of the population
did
guess randomly, we could allow some of our
agents to do that. If we were to assume that a small portion of the population
did
guess logically, we could allow some of our agents to do that. However, the remain-
der of our agents fall into that shadow zone. What we need is a way of creating the
guesses that those agents would offer. That process requires us to do more investi-
gation on the characteristics of
only
that population.
